ON THE SHAPE OF LIQUID DROPS AND CRYSTALS IN THE SMALL MASS REGIME

ON THE SHAPE OF LIQUID DROPS AND CRYSTALS IN THE SMALL MASS REGIME A. FIGALLI AND F. MAGGI Abstract. We consider liquid drops or crystals lying in equilibrium under the action of a potential energy. For small masses, the proximity of the resulting minimizers from the Wulff shape associated to the surface tension is quantitatively controlled in terms of the smallness of the mass and with respect to the natural notions of distance induced by the regularity of the Wulff shape. Stronger results are proved in the two-dimensional case. For instance, it is shown that a planar crystal undergoing the action of a small exterior force field remains convex, and admits only small translations parallel to its faces. Contents 1. Introduction 1 2. Sets of finite perimeter and volume-constrained variations 8 3. Stability properties of (ε, R)-minimizers 13 4. Stability properties of minimizers at small mass 27 Appendix A. Variation formulae and higher regularity 45 Appendix B. A remark about Question (Q1) 49 Appendix C. Almost minimal currents and C 1,α -regularity of (ε, n+1)-minimizers 50 References 57 1. Introduction 1.1. The variational problem. Let us consider a liquid drop or a crystal of mass m subject to the action of a potential. At equilibrium, its shape minimizes (under a volume constraint) the free energy, that consists of a (possibly anisotropic) interfacial surface energy plus a bulk potential energy induced by an external force field [11, 27]. Therefore one is naturally led to consider the variational problem inf { E(E) = F(E) + G(E) : E = m }. (1.1) Here, for E R n (n 2), E denotes the Lebesgue measure of E, while F(E) and G(E) are, respectively, the surface energy and the potential energy of E, that are introduced as follows. Surface energy: We are given a surface tension, that is a convex, positively 1-homogeneous function f : R n [0, + ). Correspondingly we define the surface energy of a set of finite 1

2 A. FIGALLI AND F. MAGGI perimeter E R n as F(E) = E f(ν E ) dh n 1, (1.2) where ν E is the measure theoretic outer unit normal to E and E is its reduced boundary (see section 2.1). The minimization of F under a volume constraint is described in section 1.2. Potential energy: The potential is a locally bounded Borel function g : R n [0, + ) that is coercive on R n, i.e., we have g(x) + as x +. (1.3) We also assume that inf g = g(0) = 0. (1.4) Rn This is done without loss of generality in the study of (1.1), as it amounts to subtract to the free energy a suitable constant and to translate the origin in the system of coordinates. The potential energy of E R n is then defined as G(E) = g(x) dx. (1.5) E Actually one could also allow g to take the value + in order to include a confinement constraint, since (whenever possible) a minimizer will always avoid the region {g = + }. Observe that, when g is differentiable on the (open) set {g < + }, then the energy term G(E) corresponds to the presence of the force field g acting on E. In this paper we are concerned with the geometric properties of the minimizers for the variational problem (1.1), especially in the small mass regime. Our main results are stated in Theorem 1.1 and Theorem 1.3. We remark that the coercivity assumption (1.3) excludes from our analysis the gravitational case g(x, y, z) = z in R 3 {z 0}. We use this assumption just to trivialize some side issues, such as the existence of minimizers. It is very likely that our methods could be adapted to also handle the gravitational case, and other special cases of interest, by exploiting their particular structure. 1.2. Geometric properties of Wulff shapes. In the absence of the potential term (g constant), volume-constrained minimizers of the surface energy are obtained by translation and scaling of the open, bounded convex set K known as the Wulff shape of f. The set K is explicitly given by the formula K = {x R n : (x ν) < f(ν)} = {x R n : f (x) < 1}, (1.6) ν S n 1 where we have introduced f : R n [0, + ), defined as f (x) = sup{x y : f(y) = 1}, x R n. (1.7) The minimality property of K is equivalently expressed by the Wulff inequality F(E) n K 1/n E (n 1)/n, (1.8)

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 3 where equality holds if and only if E = x + K r for some x R n. (Here and in the sequel, we use the notation K r = rk.) Indeed, the right hand side of (1.8) is equal to F(K r ) = r n 1 F(K), where r = ( E / K ) 1/n and F(K) = n K. It is useful to note that every open, bounded convex set K containing the origin is, in fact, the Wulff shape for some surface energy F = F K corresponding to the surface tension f = f K defined as f(ν) = sup{ν x : x K}. The geometric properties of a Wulff shape are closely related to the analytic properties of the corresponding surface tension. Two relevant (and somehow complementary) situations are the following ones: Uniformly elliptic case: The surface tension f is λ-elliptic, λ > 0, if f C 2 (R n \ {0}) and ( 2 f(v)τ) τ λ ) v (τ τ v v 2 v v, (1.9) whenever v, τ R n, v 0. Under these hypotheses the boundary of the Wulff shape K is of class C 2 and uniformly convex (see, for instance, [39, page 111]). Moreover, the second fundamental form ν K of K satisfies the identity 2 f(ν K (x)) ν K (x) = Id Tx K, x K. (1.10) (Notice that this makes sense as ν K (x)ν K (x) = 0 and ν K (x) ( 2 f(ν K (x))v) = 0 for every x K and v R n.) This situation includes of course the isotropic case f(ν) = λ ν (λ > 0). Evidently, in the isotropic case the Wulff shape is the Euclidean ball B λ and the Wulff inequality reduces to the Euclidean isoperimetric inequality. Isotropic (or smooth, nearly isotropic) surface energies are used to model liquid drops. Moreover, minima of the functional (1.1) with f(ν) = λ ν appear also in phase transition problems, where the mean curvature of the interface is related to the pressure or the temperature on it, represented by g (this is the so-called Gibbs-Thompson relation). Crystalline case: A surface tension f is crystalline if it is the maximum of finitely many linear functions, i.e., if there exists a finite set {x j } N j=1 Rn \ {0}, N N, such that f(ν) = max 1 j N (x j ν), ν S n 1. (1.11) The corresponding Wulff shape is a convex polyhedron. These are the surface tensions used in studying crystals [48]. 1.3. Geometric properties of minimizers. In the presence of the potential term, the geometric properties of minimizers are much less understood. A noticeable exception to this claim is the case of sessile/pendant (or otherwise constrained) liquid drops under the action of gravity. This situation, that has been extensively considered in the literature (see, for instance, [51, 20, 26]), falls into the variational problem (1.1) for the choice of an isotropic surface tension energy interacting with a potential g of the form g(x) = x n if g(x) < +. However, if we look to (1.1) in its full generality, then the validity of various natural properties of minimizers is at present unknown. In particular, the following two questions

4 A. FIGALLI AND F. MAGGI were raised by Almgren. The first question is mentioned in [32], while the second one was communicated to us by Morgan [34]. (Q1) If the potential g is convex (or, more generally, if the sub-level sets {g < t} are convex), are minimizers convex or, at least, connected? (Q2) If the surface energy dominates over the potential energy (e.g., if the potential g is almost constant or if the mass m is sufficiently small), to which extent are minimizers close to Wulff shapes? Let us point out that a question similar to (Q1) was raised by Almgren and Taylor, asking whether a crystal lying on a table under gravity is necessarily convex (see [43, Question 8] and also [10, Problem 8.4]). About the first question, in [6] the authors prove convexity of minimizers in the two-dimensional case for drops/crystals lying above a table under the action of the gravitational potential, while in [49] convexity is used as an assumption for proving (under additional suitable assumptions) facetting of a minimizing crystal. In [8, 12], by a levelsets method approach combined with convexity results for solutions to elliptic PDEs, the convexity of minimizers is proved for general convex potentials in the large mass regime. Finally, in the general planar case, it is shown in [32] that every minimizer is the union of finitely many connected components lying at mutually positive distance, all having different masses, and each component being convex and minimizing the free energy among convex sets with its same volume. With this paper, we mainly aim to stimulate the investigation of the second question, providing some optimal results, both in the planar case and in general dimension (see, however, Theorem 4.5, Theorem 4.11 and Appendix B for some results that are not related to the small mass regime). Our estimates are quantitative, in the sense that we shall present explicit bounds on the proximity to a Wulff shape in terms of the small mass m. Moreover, the value of the critical mass below, which our estimates hold can be made completely explicit from our arguments (though there will be no attempt to find such a explicit expression). Our first main result establishes the connectedness and the uniform proximity of minimizers to Wulff shapes below a critical mass. This is done for very general surface and potential energies. Theorem 1.1. There exist positive constants m c = m c (n, f, g) and C = C(n, f, g) with the following property: If E is a minimizer in the variational problem (1.1) with mass E = m m c, then E is connected and uniformly close to a Wulff shape, i.e., there exist x 0 R n and r 0 > 0, with such that where we have set r 0 C m 1/n2, x 0 + K s(m)(1 r0 ) E x 0 + K s(m)(1+r0 ), s(m) = ( ) 1/n m. K

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 5 E K Figure 1. In the small mass regime minimizers are connected and uniformly close to a (properly rescaled and translated) Wulff shape, in terms of the smallness of the mass. The convexity of these minimizers remains conjectural, with the exception of the planar case n = 2 and of the λ-elliptic case (in general dimension, see Theorem 1.3). If n = 2 then E is a convex set. Moreover, if f is crystalline (or, equivalently, if the Wulff shape K is a convex polygon), then E is a convex polygon with sides parallel to that of K. Remark 1.2. The above theorem shows that in the planar crystalline case minimizers possess a particularly rigid structure. Although our proof cannot be generalized to higher dimension, this result raises the interesting question whether or not an analogous property should hold in higher dimension (or at least in the physical case n = 3). The analogous result in higher dimension should say that, if f is crystalline, then minimizers with sufficiently small mass are polyhedra with sides parallel to that of K. This would show that a minimizer E can be obtained by K by slightly translating the faces of K, and in particular the minimization problem (1.1) would reduce to a finite dimensional problem (the dimension being equal to the number of faces of K). The main question left open by Theorem 1.1 concerns the convexity of minimizers at small mass in dimension n 3. We address this problem in the case of smooth λ- elliptic surface tensions and of potentials of class C 1. In this situation the Wulff shape turns out to be a uniformly convex set with smooth boundary. Correspondingly we prove that minimizers at small mass are not merely convex, but that they are in fact uniformly convex sets with smooth boundary and with principal curvatures uniformly close to that of a (properly rescaled) Wulff shape. To express this last property we shall make use of the second order characterization (1.10) of Wulff shapes. Theorem 1.3. If g Cloc 1 (Rn ), f C 2,α (R n \{0}) for some α (0, 1), and f is λ-elliptic, then there exist a critical mass m 0 = m 0 (n, g, f) and a constant C = C(n, g, f, α) with the following property: If E is a minimizer in (1.1) with E = m m 0 and if we set ( ) 1/n K F = E, m

6 A. FIGALLI AND F. MAGGI then F is of class C 2,α and max F 2 f(ν F ) ν F Id Tx F C m 2α/(n+2α). (1.12) In particular, if m is small enough (the smallness depending on n, f, and g only) then F (and so E) is a convex set. Remark 1.4. If g C 1,β loc (Rn ) and f C 3,β (R n \{0}) for some β > 0, then the conclusion of Theorem 1.3 can be strengthened to max F 2 f(ν F ) ν F Id Tx F C m 2/(n+2), which corresponds to (1.12) with α = 1. Remark 1.5. At a first sight, Theorem 1.3 could be seen as a slight generalization of the fact that small liquid drops lying on a table are asymptotically spherical as volume tends to zero [45], or that in a Riemannian manifold, isoperimetric regions of small volume are smoothly close to being round balls (this fact was first proved by Kleiner as explained in [50], see also [29, 38, 35]). However, our result differs from other results of this kind in the fact of being quantitative. Indeed, once uniform C 2,α -bounds for minimizers at small masses are established (see Theorem 4.6), one usually deduces their convexity by a compactness argument (indeed, as m 0, minimizers converge to K in the C 2 -topology, and the uniform convexity of K entails their convexity). In order to prove Theorem 1.3 one needs a different approach, as the use of compactness arguments rules out the possibility of finding explicit rates of convergence in terms of m 0. Observe that also the constant C appearing in (1.12) is obtained by a constructive method, and so it is a priori computable. Remark 1.6. Theorems 1.1 and 1.3 deal with the connectedness and convexity properties of liquid drops and crystals in the small mass regime. Outside this special regime, one E K Figure 2. In the planar crystalline case, minimizers are convex polygons, with sides parallel to the polygonal Wulff shape associated with the crystalline surface tension (picture on the left). The argument used in the proof of this result, when repeated in three dimensions, seems not sufficient to draw the analogous conclusions. For example, in the case of a cubic crystal, the two dimensional argument used in the proof of Theorem 1.1 allows to exclude that a cube with a rounded vertex is a minimizer, but it is not sufficient to exclude a cube with a rounded edge (see the picture on the right).

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 7 expects convexity of minimizers provided g is convex (see question (Q1)). As already mentioned, this was proved in [8, 12] when the mass is large enough. The natural problem of how to fill the gap in between these two results is open. It seems very likely that new ideas are needed to deal with this case. 1.4. Organization of the paper. In section 2 we recall some basic definitions about sets of finite perimeter, and collect some useful facts concerning surface energies and volume constrained variations. In section 3 we introduce and study the class of (ε, R)-minimizers of the surface energy F. Given ε, R > 0, a set of finite perimeter E R n is a (volume constrained) (ε, R)-minimizer of F provided for every set of finite perimeter F R n with F(E) F(F) + ε K 1/n (n 1)/n E F E, (1.13) E F = E and F I R (E), where I R (E) is the R-neighborhood of E with respect to K, i.e., I R (E) = {x R n : dist K (x, E) < R}, dist K (x, E) = inf y E f (x y). (1.14) (In the above definition, neighborhoods are defined in terms of f only to deduce cleaner estimates.) After discussing the basic regularity properties of (ε, R)-minimizers of F (Theorem 3.1), we focus on the geometric properties characteristic to the small ε regime. The L 1 -proximity (in terms of the smallness of ε) of every (ε, n+1)-minimizer to a properly rescaled and translated Wulff shape is an almost direct consequence of the main result in [19] (Theorem 3.2 and Lemma 3.3). In Theorem 3.4 and Corollary 3.5 we pave the way to the proof of Theorem 1.1 by proving that, in fact, (ε, n + 1)-minimizers are connected and uniformly close to Wulff shapes. This result may appear to the specialists as a classical application of standard density estimates combined with the above mentioned L 1 -estimate. However, at least to our knowledge, there are no universal density estimates available for (ε, R)-minimizers, i.e., density estimates independent of the minimizer. This follows from the fact that, if E is a (ε, R)-minimizer, the class of competitors has to satisfy the constraint F = E, and so we are forced to make a mass adjustments which introduces a dependence on E (see Lemma 2.3 and Theorem 3.1). For this reason the proof of Theorem 3.4, although it follows the lines of many other proofs of the same kind, presents some subtle points. This careful approach allows us to show that the uniform proximity result of Corollary 3.5 holds for every (ε, n+1)-minimizer with ε ε(n), where ε(n) depends on the dimension n only, and not on f. In section 3.3 we focus on the planar case n = 2, and show that (ε, 3)-minimizers are convex (Theorem 3.6), and that, in the crystalline case, they are convex polygons (Theorem 3.7) provided ε ε 0, where ε 0 is a universal constant independent of f. As a preparatory step towards the proof of Theorem 1.3, in section 3.4 (see also appendix C) we consider λ-elliptic surface tensions and apply the regularity theory for almost minimizing rectifiable currents to show that the

8 A. FIGALLI AND F. MAGGI boundaries of (ε, n+1)-minimizers of F satisfy uniform C 1,α -estimates for every α (0, 1) (Theorem 3.8). In section 4 we prove our main results on optimal shapes in the variational problem (1.1). The first step consists in showing that optimal shapes for (1.1) are uniformly bounded in terms of their mass, the dimension n and the way g grows at infinity. For small masses, this boundedness result is true for any potential g (Theorem 4.2). Although it is not needed for the proof of our main results, we thought it conceptually important to provide these bounds for arbitrary masses. We do this in Theorem 4.5, assuming that g is locally Lipschitz. As a by-product of the uniform boundedness result in Theorem 4.2, we immediately see that a minimizer E for (1.1) with E = m is also an (ε, n+1)-minimizer for ε C m 1/n, where C is an (explicitly computable) constant depending on n, f, and g only. This fact allows us to apply the results of section 3, to deduce Theorem 1.1 as a corollary. Moreover, when f is λ-elliptic, f C 2,α (R n \ {0}) and g C 0,α loc (Rn ) for some α (0, 1), then the regularity Theorem 3.8 for (ε, n+1)-minimizers can be combined with the first variation formula for the free energy and with elliptic regularity theory to prove that the corresponding minimizers satisfy uniform C 2,α -estimates (Theorem 4.6). Hence, in section 4.3 we apply the second variation formula for the free energy to a suitable normal vector field to show that the second fundamental form of the boundary of a minimizer E is, up to a dilation taking E into F = ( K /m) 1/n E, L 2 -close to the second fundamental form of K (Theorem 4.9). A simple interpolation between C 0,α and L 2 allows combining Theorem 4.9 with the estimates from Theorem 4.6 to show the uniform proximity of the second fundamental form of F to that of K, thus proving Theorem 1.3. Finally, in section 4.4 we use a variant of the argument in section 3.3 to show that, even outside the small mass regime, planar crystals have a remarkably rigid structure. More precisely, if f is crystalline and g is continuous then the boundary of a planar, crystalline minimizer consists of two pieces, one which is included in some level set {g = l} and the other one which is polygonal, with normal directions chosen among the normal directions to K. In appendix A we finally gather the first and second variation formulas of the free energy, together with a brief description of an useful bootstrap argument. In appendix B we make a first (small) step towards a positive answer to the convexity question (Q1), by showing that minimizers in (1.1) corresponding to potentials with convex level sets have non-negative anisotropic mean curvature (in fact, a stronger global condition is proven to hold true). Finally, appendix C reviews the regularity theory for almost minimizing currents, and shows how these kinds of results apply to our setting, to prove uniform C 1,α -regularity for (ε, n + 1)-minimizers. 2. Sets of finite perimeter and volume-constrained variations 2.1. Sets of finite perimeter. In this section we recall some basic definitions and properties on sets of finite perimeter. We refer to [4] for an extensive introduction to the subject and for a proof to all the properties stated below. A Borel set E R n is a set of

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 9 finite perimeter in R n provided { } sup div T(x) dx : T Cc 1 (R n ; B) < +, E where B = B 1 denotes the Euclidean unit ball. If this is the case, the distributional gradient D1 E of the characteristic function 1 E of E defines a Radon measure on R n, with values in R n, such that the distributional divergence theorem div T(x) dx = E T d D1 E, R n T Cc(R 1 n ; R n ), (2.1) holds true. The total variation of D1 E is then used to define the perimeter of E relative to a set A R n on setting P(E; A) = D1 E (A), P(E) = D1 E (R n ). If A is open, an equivalent definition for P(E; A), which turns out to be very useful when proving lower semicontinuity result, is also given by { } P(E; A) = sup div T(x) dx : T Cc 1 (A; B). (2.2) E If E is a bounded open set with C 1 boundary, then E is a set of finite perimeter in R n and D1 E = ν E H n 1, where ν E is the outer unit normal to E and H n 1 denotes the (n 1)-dimensional Hausdorff measure. In particular, P(E; A) = H n 1 (A ), P(E) = H n 1 (), and (2.1) amounts to the classical divergence theorem. Turning back to generic sets of finite perimeter, one see that up to modifying E on a set of measure zero (an operation that leaves D1 E unchanged) it can always be assumed that spt(d1 E ) =, [25, Proposition 3.1]. The reduced boundary E of E is then defined as the set of those x such that the limit ν E (x) = lim r 0 + D1 E (B(x, r)) D1 E (B(x, r)) exists and belongs to S n 1. It turns out that E is a countably H n 1 -rectifiable set in R n and that In particular, if A R n D1 E = ν E H n 1 E. P(E; A) = H n 1 (A E), P(E) = H n 1 ( E), and the distributional divergence theorem (2.1) takes the more appealing form div T(x) dx = T ν E d H n 1, T Cc 1 (Rn ; R n ). (2.3) E E

10 A. FIGALLI AND F. MAGGI We also recall the basic lower semicontinuity and approximation results for sets of finite loc perimeter. Let us say that E h E (resp., E h E) if 1 Eh 1 E in L 1 (R n ) (resp., L 1 loc (Rn )). If A R n loc is open and E h E then we have P(E; A) lim inf h P(E h; A). Moreover, given a set of finite perimeter E there always exists a sequence {E h } h N of open loc sets with smooth boundaries such that E h E and D1 Eh D1 E. All the relevant properties of sets of finite perimeter are left invariant by modifications on sets of (Lebesgue) measure zero. The proper notion of connectedness in this framework is then introduced as follows: a set of finite perimeter E (with finite measure) is said indecomposable if E = E 1 E 2, P(E) = P(E 1 ) + P(E 2 ), and E = E 1 + E 2 imply E 1 E 2 = 0. As a reference for indecomposable set we refer to [5]. 2.2. Basic properties of the surface energy. We now gather some basic properties of the surface energy that will be useful in the sequel. Given A R n we shall define the surface energy of the set of finite perimeter E R n relative to A as F(E; A) = f(ν E ) dh n 1, A E where, of course, F(E; R n ) = F(E). From the classical Reshetnyak theorems (see [4, Theorems 2.38-2.39], or [44] for a simpler proof) and from our basic assumptions on the surface tension f, we deduce the following lemma about the behavior of the surface energy under local convergence of sets. (The first part of the following result can be easily proven using a suitable duality formula for F as in (2.2).) Lemma 2.1. If A R n is open and E h loc E then If, moreover, P(E h ) P(E), then F(E; A) lim inf h F(E h; A). F(E; A) = lim h F(E h ; A). In proving estimates involving the surface tension f we are going to make frequent use of the quantities 0 < α 1 α 2 < + defined as In particular, α 1 = min S n 1 f, α 2 = max f. (2.4) Sn 1 α 1 f L (R n ;R n ) α 2. (2.5) Indeed, by the positive 1-homogeneity and by the convexity of f we immediately see that f is sub-additive, so that f(x + t v) f(x) f(t v) tα 2

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 11 for every x R n, t > 0, and v S n 1. On the other hand, if f is differentiable at x R n then f(x) (x/ x ) = f(x/ x ) α 1. Thus, α 1 sup{ f(x) v : v S n 1 } = f(x) α 2, and (2.5) immediately follows. It is also useful to note that the dual function f to f introduced in (1.7), that is still convex and positively 1-homogeneous, satisfies inf f = 1, S n 1 α 2 sup f = 1, (2.6) S n 1 α 1 1 α 2 f L (R n ;R n ) 1 α 1. (2.7) Eventually we notice from (1.6) that B α1 is the largest Euclidean ball centered at the origin that is contained in K, while B α2 is the smallest Euclidean ball centered at the origin that contains K. Thus, α 1 = sup{r > 0 : B r K}, α 2 = inf{r > 0 : K B r }. (2.8) Because of (2.4), (2.5), (2.6), (2.7), and (2.8) we shall often produce estimates depending on the ratio α 2 /α 1. One is usually able to rule out such a dependence by means of the following lemma. Lemma 2.2 (A normalization lemma). If K is an open, bounded convex set with 0 K, then there exist an affine map L : R n R n with det L = 1 and r = r(n, K ) > 0, such that If f L(K) is the surface tension associated to L(K), then B r L(K) B n r. (2.9) sup S n 1 f L(K) inf S n 1 f L(K) n, F K (E) = F L(K) (L(E)), (2.10) for every set of finite perimeter E R n. Moreover, if g is a bounded Borel function and we set g L = g L 1, then g = g L. (2.11) E Proof. By John s Lemma [28, Theorem III], we may associate to K an affine map L 0 : R n R n such that det L 0 > 0 and B 1 L 0 (K) B n. Therefore, up to the multiplication of L by a constant, we can achieve (2.9). Clearly, (2.11) is a trivial consequence of the fact that det L = 1. Finally, to show (2.10) let us now recall that, if E is a bounded open set with smooth boundary, then L(E) E + εk E F K (E) = lim. ε 0 + ε Since L is affine with det L = 1, L(E) + εl(k) L(E) = L(E + εk) E = E + εk E, hence F K (E) = F L(K) (L(E)). By a density argument and Lemma 2.1 we immediately get (2.10).

12 A. FIGALLI AND F. MAGGI 2.3. Volume-constrained variations. In studying minimizers to the variational problem 1.1, we will often construct suitable comparison sets, that are typically obtained by a cut and paste operation followed by a mass adjustment (usually done by a dilation). Having in mind to work with the notion of (ε, R)-minimizers introduced in (1.13), we have to be careful to control the surface energy variation and the L 1 -distance variation created in the mass adjustment. We will adjust mass in two ways: either by a first variation argument (where the surface energy variation depends on the set itself in a quite involved way), see Lemma 2.3 (this lemma is sometimes referred to as Almgren s Lemma, see [36, Lemma 13.5]); or by a scaling operation (in this case the surface energy variation is trivial but the L 1 -distance variation requires an estimate), see Lemma 2.4. We now prove these technical lemmas. Lemma 2.3. If E is a set of finite perimeter in R n and A is an open set such that A E is nonempty, then there exist s 0 = s 0 (E, A) and C = C(E, A, α 2 ) such that, for every s ( s 0, s 0 ), there exists a set of finite perimeter F with the following properties: E F A, E F = s, F(E) F(F) C s. Proof. Let T Cc (A; Rn ) and Φ t (x) = x + t T(x), x R n. There exists t 0 > 0 such that Φ t is a diffeomorphism of R n whenever t < t 0. Hence Φ t (E) is a set of finite perimeter for every t < t 0, with Φ t (E) E A. By the first variation formulae in Appendix A.1 we have Φ t (E) = E + t T ν E dh n 1 + O(t 2 ), E F(Φ t (E)) = (1 + t div T + O(t 2 )) f(ν E t( T) ν E + O(t 2 )) dh n 1, (2.12) E where, here and in the rest of the proof, we denote by O(s) a function of s such that O(s) C s for a constant C depending on T only. In order to estimate F(Φ t (E)) F(E) we now notice that by (2.5) f(ν E t( T) ν E + O(t 2 )) f(ν E ) α 2 ( t T + O(t)), (2.13) while, thanks to (2.4) and the simple inequality we also have ν E t( T) ν E + O(t 2 ) 1 + t T + O(t 2 ), t div T + O(t 2 ) f(ν E t( T) ν E + O(t 2 )) α 2 ( t T + O(t 2 )). (2.14) By combining (2.12), (2.13), and (2.14), we find that ) F(E) F(Φ t (E)) 2α 2 ( t T dh n 1 + O(t 2 ). E

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 13 Since A E is non-empty, it is easily seen that there exists T Cc (A; R n ) such that γ = T ν E dh n 1 > 0, E for instance on setting T = ϕ ν E (x 0 ) for x 0 A E and ϕ Cc (A) such that 1 B(x,r/2) ϕ 1 B(x,r) (with r sufficiently small). Therefore, the function t Φ t (E) = E + tγ + O(t 2 ) is injective on some open interval ( t 0, t 0 ) where γ 2 t E Φ t(e), F(E) F(Φ t (F)) 4α 2 t T dh n 1. We conclude by choosing s 0 > 0 such that the interval ( E s 0, E + s 0 ) is contained in the image of ( t 0, t 0 ) through t Φ t (E), proving the result with the constant C defined as C = 8 α 2 T dh n 1. γ E Lemma 2.4. There exists a constant C(n) with the following property: If E is a set of finite perimeter with E B R, then whenever λ (1/2, 2). E E (λ E) C(n) λ 1 R P(E), (2.15) Proof. If u C 1 c (Rn ) and λ [1/2, 2], then for every x R n u(x) u(x/λ) 2 λ 1 x 1 0 u(x + t(1 1/λ)x) dt. If spt(u) B R, then by Fubini theorem we have 1 u(x) u(x/λ) dx λ 1 R dt u(x + t(1 1/λ)x) dx R n 0 R ( n 1 ) dt λ 1 R u 0 (1 + t(1 1/λ)) n R n C(n) λ 1 R u. R n We prove (2.15) by testing this inequality on u ε = 1 E ρ ε and letting ε 0 +. 3. Stability properties of (ε, R)-minimizers 3.1. Basic properties of (ε, R)-minimizers. This section is devoted to the study of geometric properties of (ε, R)-minimizers. Given ε, R > 0, let us recall that a set of finite perimeter E R n is a (volume constrained) (ε, R)-minimizer of the surface energy F = F K provided F(E) F(F) + ε K 1/n (n 1)/n E F E, E

14 A. FIGALLI AND F. MAGGI for every set of finite perimeter F R n with F = E and F I R (E), where I R (E) is defined as the R-neighborhood of E with respect to f, see (1.14). Note that (ε, R)- minimality is formulated to be a scale invariant property with respect to ε. Indeed, if E is an (ε, R)-minimizer of F K, then for every λ > 0 the rescaled set λ E is an (ε, λ R)- minimizer of F K, or equivalently an (ε, R)-minimizer of F λk. More in general, if L is an affine transformation with det L > 0, Lemma 2.2 and the discussion above gives that L(E) is a (ε, R)-minimizer of F L(K). Of course, the Wulff shape K is an (ε, R)-minimizer of F K for every ε and R. Theorem 3.1 (Basic regularity estimates for (ε, R)-minimizers). If E is an (ε, R)- minimizer of F then H n 1 ( \ E) = 0 and E is equivalent to its interior. Moreover, is differentiable at every point of E, i.e., lim sup r 0 + for every x 0 E. { } (x x0 ) ν E (x 0 ) : x B(x 0, r), x x 0 = 0, (3.1) x x 0 Proof. Step one. As stated in section 2.1, up to modifying E on a set of measure zero, we can assume that = spt(d1 E ) = {x R n : 0 < E B(x, r) < ω n r n for every r > 0}. We now claim that there exist positive constants κ = κ(n, f, E, R) < 1 and r = r(n, f, E, R), such that if x and r < r, then H n 1 (B(x, r) E) κ r n 1, (3.2) ω n r n (1 κ) B(x, r) E κω n r n. (3.3) Of course, it will suffice to show that these estimates hold for a.e. r < r. Hence, thanks to the coarea formula [4, Theorem 2.93] applied with the Lipschitz function x x and the countably H n 1 -rectifiable set E, we may restrict to consider values of r < r such that H n 1 ( B(x, r) E) = 0. (3.4) Fix x 1 x 2, and let r 0 = r 0 (E) > 0 be such that 2r 0 < α 1 R, B(x 1, 2r 0 ) B(x 2, 2r 0 ) =, (recall that, by definition (2.4) of α 1, we have B α1 R K R ). Let s 1, s 2 > 0 and C 1, C 2 be the constants given by Lemma 2.3 applied to E on the open sets B(x 1, r 0 ) and B(x 2, r 0 ) respectively, and notice that s k < B(x k, r 0 ) = ω n r0 n. We set, and require r to satisfy s = min{s 1, s 2 }, C = max{c 1, C 2 }, ω n r n < s, so that, in particular, r < r 0. Since the balls B(x 1, 2r 0 ) and B(x 2, 2r 0 ) are disjoint, we can decompose as M 1 M 2, where M k = {x : B(x, r) B(x k, r 0 ) = }, k = 1, 2,

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 15 so that \ B(x k, 2r 0 ) M k. We are now in the position to prove (3.2) and (3.3). Let x M 1 and consider the function u(r) = E B(x, r) = r 0 H n 1 (E B(x, s)) ds. If we set G = E \ B(x, r) for some r < r such that (3.4) holds true, then we have 0 < E G = u(r) < ω n r n < s. Since E B(x 2, r 0 ) = G B(x 2, r 0 ), we can apply Lemma 2.3 to find a set of finite perimeter F such that F G B(x 2, r 0 ), F = G + ( E G ) = E, and F(F) F(G) + C G F = F(G) + C ( E G ) = F(G) + Cu(r). We can now test the (ε, R)-minimality of E against F to find 1/n E F u(r) F(E) F(G) + Cu(r) + ε K = F(G) + Cu(r) + 2ε K 1/n E 1/n E, 1/n where we used that by construction E F = 2u(r). Moreover, by (3.4), we have that F(E) F(G) = f(ν E ) dh n 1 f(ν B(x,r) ) dh n 1 B(x,r) E B(x,r) E 1 Hence we get α 1 H n 1 (B(x, r) E) α 2 u (r). P(E; B(x, r)) C(u (r) + u(r)) for some constant C = C(E, f). Since P(E B(x, r)) = P(E; B(x, r)) + u (r), due to the Euclidean isoperimetric inequality, we also have nω 1/n n u(r)(n 1)/n P(E B(x, r)) C(u (r) + u(r)), (3.5) for every r < r such that (3.4) holds. Since u(r) ω n r n, up to further decrease r (depending on n and on the constant C appearing in (3.5)), we may assume that Cu(r) = Cu(r) (n 1)/n u(r) 1/n nω1/n n u(r) (n 1)/n, 2 whenever r < r, so that (3.5) implies c u(r) (n 1)/n u (r), for a.e. r < r and for some c = c(n, f, E, R). Since u(r) > 0 for every r > 0 we deduce that (u(r) 1/n ) c for a.e. r < r. Hence u(r) c r n for r < r, that is the lower bound in (3.3). The upper bound in (3.3) follows by an entirely similar argument, where we consider G = E B(x, r) instead of G = E \ B(x, r). This remark completes the proof of (3.3). Now that (3.3) has been proved we can apply the relative isoperimetric inequality in B(x, r) (see, for instance, [4, Remark 3.45]) to see that H n 1 (B(x, r) E) τ(n) min{ E B(x, r), B(x, r) \ E } (n 1)/n τ(n)cr n 1, and (3.2) is proved.

16 A. FIGALLI AND F. MAGGI Step two. We now conclude the proof of the theorem. First of all, (3.2) combined with a standard covering argument (see, e.g., [3, Corollario 4.2.4]) implies that H n 1 ( \ E) = 0, hence that E is equivalent to its interior. To show the differentiability of at the points of E, let us recall that if x 0 E then (see, e.g., [4, Theorem 3.59]) E x0,r = E x 0 r loc {x R n : ν E (x 0 ) x 0} as r 0 + (recall that the above convergence of sets means that their characteristic functions converge in L 1 loc ). We now show that this L1 loc-convergence combined with the density estimate (3.3) implies that convergence holds in the Hausdorff sence, which is actually equivalent to (3.1). Indeed, for every σ > 0 we have Ex0 lim,r {x B : x ν E (x 0 ) σ} = {x B : 0 x ν E (x 0 ) σ} ω n 1 σ, r 0 + so there exists r σ = r σ (E, x 0 ) > 0 such that {x E B(x 0, r) : (x x 0 ) ν E (x 0 ) σr} 2ω n 1 σr n, (3.6) for every r < r σ. Let us now chose L = L(n, f, E, R) > 0 such that κl n > 2ω n 1, (3.7) where κ is the constant found in Step one. We claim that if σ > 0 is small enough with respect to L, then for every r < r σ we have {x B(x 0, r/2) : (x x 0 ) ν E (x 0 ) Lσ 1/n r} =. (3.8) Indeed if (3.8) is not true, then there exists x 1 such that, provided σ is small enough, E B(x 1, Lσ 1/n r) {x E B(x 0, r) : ν E (x x 0 ) σr}, E B(x 1, Lσ 1/n r) κl n σr n, where we have also taken (3.3) into account. Since the combination of these two facts would lead to a contradiction with (3.6) and (3.7), we conclude that (3.8) holds true. By an analogous argument one proves that, for the same values of L and σ, {x B(x 0, r/2) : (x x 0 ) ν E (x 0 ) Lσ 1/n r} =, (3.9) whenever r < r σ. On combining (3.8) and (3.9) we conclude that for every σ > 0 there exists r σ > 0 such that (x x 0 ) ν E (x 0 ) Lσ 1/n x x 0, for every x B(x 0, r σ /2), so that (3.1) is proved.

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 17 3.2. Stability properties of (ε, R)-minimizers at small ε. Given a set of finite perimeter E, we define its Wulff deficit (with respect to F) as δ(e) = F(E) 1. (3.10) n K 1/n E (n 1)/n By the Wulff inequality (1.8) we have δ(e) 0, and the characterization of the equality case gives that δ(e) = 0 if and only if E (x + K r ) = 0 for some x R n and r > 0. In [19] we have proved the following theorem, that gives a (sharp) strengthened form of the Wulff inequality (1.8). Theorem 3.2. If E is a set of finite perimeter in R n with E = K then there exists x 0 R n such that ( ) 2 E (x0 + K) δ(e) C(n), (3.11) K or, equivalently, ( ) } 2 F(E) n K 1/n E (x0 + K) E {1 1/n + C(n), K where C(n) is a constant depending on the dimension n only. The above result says that, if δ(e) is small, then E is close in L 1 -norm to a translation of K. As the next lemma shows, (ε, n + 1)-minimizers have small deficit for ε sufficiently small (the choice R = n + 1 comes from the fact that we need all the translations x + K with E (x + K) to be admissible competitors). Lemma 3.3. If E is an (ε, n + 1)-minimizer of F with E = K then Moreover, there exists x 0 R n such that δ(e) C(n) ε 2. (3.12) E (x 0 + K) C(n) K ε, (3.13) where C(n) is a constant depending on the dimension n only. Proof. Step one. If L is as in Lemma 2.2, then L(E) is an (ε, n + 1)-minimizer of F L(K), with δ(e) = δ L(K) (L(E)), E = L(E), K = L(K) and E (x+k) = L(E) (L(x)+ L(K)) for every x R n. Therefore we may assume without loss of generality that K satisfies B r K B rn, for some r = r(n, K ). In particular, α 2 /α 1 n (see (2.4)). Step two. If E (x + K), then x + K I n+1 (E). Indeed, given z E (x + K), y x + K, and taking into account that K = {f < 1}, then we have f (y z) f (y x) + f (x z) 1 + α 2 α 1 f (z x) 1 + n.

18 A. FIGALLI AND F. MAGGI Therefore, if x R n is such that E (x + K), then by the (ε, n + 1)-minimality of E we find F(E) F(x + K) + ε E (x + K) = F(K) + ε E (x + K). Since F(K) = n K and E = K, this implies δ(e) = F(E) n K 1 E (x + K) n K On choosing x = x 0 such that (3.11) holds, we deduce (3.13). Next, by inserting (3.13) in the above estimate, we also find (3.12). We now prove an uniform stability estimate together with a connectedness results. In order to apply this result in the case of minimizers to the variational problem (1.1) we work with a slightly different notion of minimality rather than (ε, R)-minimality. The theorem is then applied to (ε, R)-minimizers in Corollary 3.5 below. Theorem 3.4. There exist constants C(n) and ε(n) with the following property: If 0 < ε < ε(n) and if E is a set of finite perimeter with E = K, such that ε. δ(e) C(n)ε 2, (3.14) E K C(n) K ε, (3.15) F(E) F(F) + ε E F, (3.16) whenever F = E, F \ E K 3, then E is indecomposable and for some r 0 C(n)ε 1/n, K 1 r0 E K 1+r0. Corollary 3.5 (Uniform proximity to the Wulff shape). There exist constants C(n) and ε(n) with the following property: If E is an (ε, n + 1)-minimizer of F with E = K and ε < ε(n), then E is connected and there exists x 0 R n and r 0 C(n)ε 1/n such that x 0 + K 1 r0 E x 0 + K 1+r0. (3.17) Proof of Corollary 3.5. It follows immediately from Theorem 3.1, Lemma 3.3, Theorem 3.4, and the fact that an open indecomposable set with H n 1 ( \ E) = 0 is connected [5, Theorem 2]. Proof of Theorem 3.4. We can apply Lemma 2.2 and assume without loss of generality that, for a constant ρ = ρ(n, K ), we have ρ α 1 α 2 n ρ (see (2.4)). Since ρ 1 E satisfies the minimality condition (3.16) with F ρ 1 K and ρ 1 K in place of F and K, up to scale both E and K by the factor 1/ρ, we may work under the additional assumptions that B K B n, (3.18) 1 α 1 α 2 n, (3.19) 1 n f L (R n ;R n ) 1, (3.20)

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 19 where (3.20) follows from (3.19) and (2.7). We notice that by (3.14) and by taking into account that F(K) = n K n B n C(n) we have F(E) ( 1 + C(n)ε 2) F(K) C(n) K C(n). (3.21) By [18, 4.2.25], [5, Theorem 1], there are countably many disjoint sets of finite perimeter {E h } h N such that E = E h, P(E) = P(E h ), h N h N and each E h is indecomposable, in the sense that if F E h is a set of finite perimeter with P(E h ) = P(F) + P(E h \ F) then F E h \ F = 0. In fact, the reduced boundaries of the E h s are pairwise disjoint mod-h n 1, so that we also have F(E) = h N F(E h ). (3.22) Without loss of generality we may assume that E 1 E h for every h N. Step one: L 1 -estimates for E 1. We claim that E \ E 1 C(n)δ(E) n, (3.23) E 1 K C(n)ε. (3.24) Since E 1 K E K + E \ E 1, (3.24) is an immediate consequence of (3.15), (3.23), and (3.14). Hence we directly focus on the proof of (3.23). Without loss of generality we may assume that E h for some h > 1. Then for every k 1 we introduce the sets of finite perimeter k F k = E h, G k = E h. h=1 h=k+1 By (3.22), by the Wulff inequality and by concavity of t 1/n on t > 0, we find that F(E) = F(F k ) + F(G k ) n K 1/n ( F k 1/n + G k 1/n ) n K 1/n E 1/n, i.e., by the definition of δ(e) (3.10), ( Fk δ(e) E ) 1/n + ( 1 F ) 1/n k 1. E Observe now that there exists a constant c 0 (n) > 0 such that t 1/n + (1 t) 1/n 1 c 0 (n)t 1/n, t [0, 1/2]. Hence, if E 1 E /2 and we chose k = 1, then we find ( δ(e) c 0 1 E ) 1/n 1 E that is (3.23) as required. Let us now assume on the contrary that E 1 < E /2, then there exists k 2 such that F k 1 < E 2, G k 1 E 2, F k E 2, G k < E 2.,

20 A. FIGALLI AND F. MAGGI Therefore, by the above argument we deduce that so that δ(e) c 0 ( Fk 1 E ) 1/n, δ(e) c 0 ( Gk E ) 1/n C(n)δ(E) n F k 1 + G k = 1 E k E E 1 E 1 E 1 2. Hence, if E 1 < E /2, then δ(e) δ(n) > 0, and this can be excluded by (3.14) provided ε ε(n) for ε(n) small enough. Step two: Uniform outer estimate for E 1. We now prove that, if ε(n) is sufficiently small, then there exists r (1, 1 + C(n)ε 1/n ) such that E 1 K r. We consider the decreasing function u : [0, + ) [0, + ) defined as u(r) = f (x) dx, r > 0. E 1 \K r By the coarea formula u is absolutely continuous, with u(r) = + and moreover (using (3.20)) We define and prove that r H n 1 (( K s ) E 1 ) ds, for every r > 0, u (r) = H n 1 (( K r ) E 1 ), for a.e. r > 0, 1 n E 1 \ K r u(r) E 1 \ K r. (3.25) r 1 = sup{r 1 : u(r) > 0}, r 1 1 C(n)ε 1/n. (3.26) For every r (1, min{r 1, 2}) we have (E 1 K r ) (E \ E 1 ) < E, therefore we can find s = s(r) > 1 such that F = E, where we have defined Set Observe that, by (3.25) and by (3.24), while, (3.23) and (3.14) give By the definition of F we have that F = s(e 1 K r ) (E \ E 1 ). v(r) = s(e 1 K r ) (E \ E 1 ), r > 1. u(r) E 1 \ K C(n)ε, (3.27) v(r) E \ E 1 C(n)ε 2n C(n)ε. (3.28) E = F = E \ E 1 + s(e 1 K r ) s(e 1 K r ) (E \ E 1 ),

which by (3.25) gives SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 21 = E \ E 1 + s n ( E 1 E 1 \ K r ) v(r), E 1 + v(r) s n ( E 1 n u(r)). Taking (3.23) and (3.27) into account, we find that 1 < s 1 + C(n) ( u(r) + v(r) ). (3.29) Hence by a suitable choice of ε(n) and by taking (3.18) into account, we can bound s through (3.29) so to entail F \E K 3. Observing that E F = 2 F \E, the minimality condition (3.16) implies On the one hand, we remark that F(E) = F(E 1 ) + F(E \ E 1 ) F(E) F(F) + 2ε s (K r E 1 ) \ E 1. (3.30) F(F) F(E \ E 1 ) + s n 1 F(K r E 1 ) F(s(E 1 K r ) (E \ E 1 )), so that (3.30) gives F(E 1 ) + F(s(E 1 K r ) (E \ E 1 )) s n 1 F(K r E 1 ) + 2ε s (K r E 1 ) \ E 1. (3.31) On the other hand, as r (1, min{r 1, 2}) we have K r B 2n, and by Lemma 2.4 we find s (K r E 1 ) \ E 1 s(k r E 1 ) \ (K r E 1 ) C(n)(s 1)F(K r E 1 ). (3.32) We now combine (3.31) and (3.32): taking also (3.29) into account and applying the Wulff inequality to s(e 1 K r ) (E \ E 1 ), ( F(E 1 ) + n K 1/n v(r) 1/n 1 + C(n) [ u(r) + v(r) ]) F(K r E 1 ). (3.33) We notice that for a.e. r > 0 F(E 1 ) = f(ν E ) dh n 1 + f(ν E ) dh n 1, (3.34) ( 1 )\K r ( 1 ) K r F(K r E 1 ) = f(ν E ) dh n 1 + f(ν Kr ) dh n 1 ( 1 ) K r ( K r) E 1 f(ν E ) dh n 1 + n u (r), (3.35) ( 1 ) K r (as α 2 n). Moreover, as r (1, min{r 1, 2}), by (3.21) F(E 1 K r ) F(E) + F(K r ) C(n). We combine this last estimate with (3.33), (3.34), and (3.35) to obtain f(ν E ) dh n 1 + c(n)v(r) 1/n n u (r) + C(n)(u(r) + v(r)). (3.36) ( 1 )\K r By applying Wulff s inequality on E 1 \ K r, thanks to (3.18) and (3.19) we find c(n)u(r) 1/n n K 1/n E 1 \ K r 1/n F(E 1 \ K r ) = f(ν E ) dh n 1 + f( ν Kr ) dh n 1 ( 1 )\K r ( K r)\e 1

22 A. FIGALLI AND F. MAGGI ( 1 )\K r f(ν E ) dh n 1 + n u (r), that combined with (3.36) leads to { u(r) 1/n + v(r) 1/n C(n) u (r) + ( u(r) + v(r) )}, (3.37) for every r (1, min{r 1, 2}). By (3.27) and (3.28) we can chose ε(n) small enough to ensure that v(r) 1/n C(n)v(r), u(r) 1/n 2 C(n) u(r), where C(n) is the constant appearing on the right hand side of (3.37). As a consequence, for every r (1, min{r 1, 2}). Thus, min{r 1, 2} 1 C(n) u(r) 1/n C(n) u (r), min{r1,2} 1 C(n)u(1) 1/n C(n)ε 1/n, u (r) u(r) 1/n dr = C(n) ( u(1) 1/n u(min{r 1, 2}) 1/n) where in the last step we have applied (3.15). Hence, if ε(n) is chosen sufficiently small, this last estimate implies r 1 1 + C(n)ε 1/n, that is (3.26), as required. Step three: Inner estimate. We now set and show that r 0 = sup{r [0, 1] : K r \ E = 0}, 1 r 0 C(n)ε 1/n. (3.38) To this end we notice that for every r (r 0, 1) we have (E 1 K r ) (E \E 1 ) > E. Thus s = s(r) (0, 1) can be defined with the property that, if we set F = s(e 1 K r ) (E \ E 1 ), then F = E. Since we have proved in Step two that E 1 K 2, we clearly have F \ E K 3. Hence we can exploit the minimality condition (3.16) to compare E and F, and deduce by the very same argument used in Step two that u(r) 1/n C(n)u (r), for a.e. r (r 0, 1). Of course, now u : [0, + ) [0, + ), is the absolutely continuous, increasing function defined as r u(r) = f (x) dx = H n 1 (( K s ) \ E) ds, r > 0. K r\e We leave the details to the interested reader. 0 Step four: E is indecomposable. We are going to prove that E = E 1. Indeed let us now set F = s E 1 where s n E 1 = E. Recalling that E 1 K 2, since E \ E 1 C(n)δ(E) n C(n)ε 2n and 1 s 1 + C(n) E \ E 1, (3.39)

SHAPES OF LIQUID DROPS AND CRYSTALS AT SMALL MASSES 23 if ε(n) is small enough then we clearly have F \ E K 3. By the minimality condition (3.16), so that by (3.39) F(E 1 ) + F(E \ E 1 ) = F(E) s n 1 F(E 1 ) + ε (s E 1 ) E, F(E \ E 1 ) = (s n 1 1)F(E 1 ) + ε (s E 1 ) E C(n) E \ E 1 F(E 1 ) + ε (s E 1 ) E. As F(E 1 ) F(E) C(n) (see (3.21)), by applying the Wulff inequality to E \ E 1 and taking into account that we have (s E 1 ) E (s E 1 ) E 1 + E \ E 1, n K 1/n E \ E 1 1/n C(n) E \ E 1 + ε (s E 1 ) E 1. By Lemma 2.4 and since E 1 K 2 B 2n, Hence, (s E 1 ) E 1 C(n)(s 1)P(E 1 ) C(n)(s 1) C(n) E \ E 1. E \ E 1 1/n C(n) E \ E 1. By (3.23) this is impossible for ε(n) small enough, unless E \ E 1 = 0. 3.3. Geometric properties of planar (ε, R)-minimizers at small ε. In this section we restrict our analysis to the planar case n = 2. This allows us to take advantage of the fact that the surface energy F does not increase under convexification to show some strong stability results of (ε, R)-minimizers. Indeed, we will prove that (ε, R)-minimizers are always convex for ε small enough (Theorem 3.6). Moreover, if the surface tension is crystalline, then (ε, R)-minimizers enjoy exactly the same crystalline structure of K (see Theorem 3.7 below). Theorem 3.6 (Convexity of planar (ε, 3)-minimizers at small ε). Let n = 2. There exists a positive constant ε 0 > 0 such that, if E is an (ε, 3)-minimizer of F with E = K and ε ε 0, then E is convex. Proof. As in the proof of Theorem 3.4, we can assume without loss of generality that B 1 K B 2. By that theorem, provided ε 0 is small enough and up to a translation, we also know that K 1 r0 E K 1+r0, where r 0 C ε 1/2. Let now F = co(e) denote the convex hull of E, and assume by contradiction that δ = F \ E / E > 0. Since, by construction, K 1 r0 F K 1+r0, we find that δ = F \ E K Therefore, if we rescale F and define K 1+r 0 \ K 1 r0 K F = (1 + δ) 1/2 F, = 2r 0 C ε 1/2. (3.40)

24 A. FIGALLI AND F. MAGGI then F = E and, provided ε 0 is small enough, F I 3 (E). Since F is obtained by a contraction of the convex set F with respect to 0 F, we have F F. Hence, E \ F F \ F = F F = δ F = δ E, 1 + δ F \ E F \ E = δ E. Moreover, as E R 2, the convexity of f ensures that F(F) F(E) [32, Corollary 2.8]. In conclusion, the (ε, 3)-minimality of E implies that F(E) F(F ) + ε E F F(E) + 2ε δ E (1 + δ) 1/2 (1 δ2 ) + o(δ) F(E) + 2ε δ E. By the Wulff inequality (1.8), F(E) 2 E, hence δ 2 + o(δ) ε δ, which combined with (3.40) leads to a contradiction for ε 0 small enough. Theorem 3.7 (Crystalline structure of (ε, 3)-minimizers at small ε). Let n = 2 and let f be a crystalline surface tension, so that the Wulff shape K is a convex polygon with outer unit normals {ν i } N i=1. There exists a positive constant ε 0 such that, if E is an (ε, 3)-minimizer with ε ε 0, then E is a convex polygon with ν E (x) {ν i } N i=1 for H 1 -a.e. x. Proof. Every affine transformation L maps a convex polygon K into a convex polygon L(K), and an (ε, 3)-minimizer E of F K into an (ε, 3)-minimizer L(E) of F L(K). Hence, thanks to Lemma 2.2 and up to a dilation, we can assume without loss of generality that B 1 K B 2. In particular, provided ε 0 is sufficiently small and up to a translation, we can apply Corollary 3.5 to entail E B 2(1+r0 ) B 4. (3.41) We now order the normal directions {ν i } N i=1 in the clockwise direction (see Figure 3), and write, with abuse of notation, ν 1 < ν 2 <... < ν N < ν 1 = ν N+1. For every j {1,..., N}, we see from (1.11) that there exists x j R n \ {0} such that f(ν) = x j ν, if ν j ν ν j+1. (3.42) If ε 0 is small enough, then by Theorem 3.6 E is a convex open set. In particular, there exists a continuous injective function γ : [0, 1) R 2 with γ(1 ) = γ(0) and γ([0, 1)) =, such that γ is of class C 1 over J = (0, 1)\I, where I consists of at most countably many points. Correspondingly, the classical outer unit normal ν E is defined as a continuous vector field over γ(j). We now claim that if t 0 J, x 0 = γ(t 0 ), then ν E (x 0 ) {ν i } N i=1. We can argue by contradiction, assuming on the contrary (and without loss of